Journal of Economic Behavior & Organization Vol.

61 (2006) 668690

Organizational performance in hierarchies and communities of practice

Olivier Dupou et a , Murat Yldzo glu b,
a

BETA (UMR CNRS 752) Universit e Louis Pasteur, 61, Avenue de la For et Noire,67000 Strasbourg, France b E3i, IFReDE-GRES, Universit e Montesquieu Bordeaux IV, Avenue L eon Duguit, 33608 Pessac, France Received 26 October 2003; accepted 21 July 2004 Available online 24 July 2006

Abstract In an earlier article we studied the Communities of Practice and their conditions of emergence using an Agent based model with a set of agents facing a continuous ow of problems. We now center our analysis on the performance of this organizational structure compared to a two-level hierarchical delegation structure. Our results show the crucial role played by the communication and the specialisation of agents; especially that community structures are efcient for competence building and learning in the long term. This paper backs the claim made by (Bowles, S., Gintis, H., 2002. Social capital and community governance. The Economic Journal 112, 419437.) that hierarchy and communities are complementary modes of governance. 2006 Elsevier B.V. All rights reserved.
JEL classication: D2; D83; L2 Keywords: Communities of practice; Learning; Emergence of networks; Organisational efciency; Hierarchy

1. Introduction The idea that informal social structures play an important role in the behavior and capabilities of organizations is a long-standing idea in disciplines such as the sociology of organizations (Crozier and Friedberg, 1977) or management studies (Brown and Duguid, 1991, 1998). The studies of such informal structures have attained a growing interest with the recent emphasis put on the knowledge-based economy and the collective learning processes that become the key factor for success of rms in such a framework. Indeed, an important part of the learning processes within organizations are increasingly seen as taking place within the informal networks nested in them.

Corresponding author. Tel.: +33 556 84 54 53; fax: +33 556 84 86 47. E-mail addresses: dupouet@cournot.u-strasbg.fr (O. Dupou et), yildi@u-bordeaux4.fr (M. Yldzo glu).

0167-2681/$ see front matter 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.jebo.2004.07.011

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

669

In particular, the concept of community of practice is deemed particularly useful by a number of scholars to account for the learning processes taking place within an organization (Brown and Duguid, 1991, 1998; Wenger, 1998). The concept of community of practice was introduced by Lave and Wenger (1990) who, by focusing on individuals practices, identied groups of persons engaged in the same practice communicating regularly with one another about their activities. Members of a community of practice essentially seek to develop their competencies in the considered practice. Communities of practice can then be seen as a means to enhance individual competencies; they are oriented toward their members (Lave and Wenger, 1990; Brown and Duguid, 1991). This goal is reached through the construction, the exchange and the sharing of a common repertoire of resources (Wenger). According to this denition, communities of practice appear as important loci for competence building within rms. As Wenger puts it, communities of practice are elementary regimes of competence. Amin and Cohendet (2003, 74) note These communities might be found in traditional work divisions and departments, but they also cut across functional divisions, spill over into after-work or project-based teams, and straddle networks of cross-corporate and professional ties. For example, within rms, classical communities include functional groups of employees who share a particular specialisation corresponding to the classical division of labor (e.g. marketing or accounting). They also include teams of employees with heterogeneous skills and qualications, often coordinated by team leaders and put together to achieve a particular goal in a given period of time. However, although some works, either empirical or theoretical, investigate the impact of communities of practice upon the performance of the organization as a whole (Orr, 1990; Huberman and Hogg, 1995), such studies remain rare. More precisely, some measures of performance of communities themselves would be desirable. Moreover, the interplay between communities and hierarchy and the output of this interplay deserves deeper analysis. These issues are difcult to tackle in traditional ways since the impact of communities upon hierarchical structures and their potential benets in terms of performance remain largely invisible in case studies. Their existence can be evidenced, but the true input of their activities for the rm are hard to assess. Besides, most of the theoretical work on communities of practice leaves some questions (such as the boundary of communities, for instance) unanswered. The concept is thus difcult to formalize using, for example, a mathematical apparatus. In a previous work (Dupou et et al., 2003), we explored the emergence of communities of practice from a collection of individuals engaged in problemsolving activities. We evidenced sufcient conditions for the emergence of such social structures and show the usefulness of some indicators to characterize networks of communities of practice. This paper goes one step further and, using computational methods, seeks to shed some light on the role of communities in the performance of organizations, using computational methods. Even if this articles main contribution concerns the emergence and the efciency of the communities of practice, we use the hierarchy as a benchmark case. As a consequence this article is somewhat connected to the literature that has been mainly developed following the seminal article of Bolton and Dewatripoint (1994). This is not the place to review this quite rich literature, but we can emphasize the complementarity that exists between our work and these articles on the efciency of hierarchy. Bolton and Dewatripoint consider the efciency of different structures for the communication networks in an organization that must handle a continuous ow of information. They study the trade-off that must be established between the specialisation of the agents and the cost of communications under differentgivennetwork structures and characterize the conditions under which a pyramidal structure can become the most efcient one. Even if the main problem of this article is not very far from some of the points we study in our model, the structure of these models are very different. In our case the structure of the communication

670

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

network is not given, but it results from the interactions between agents. We show that under some conditions, this structure can take the form of a communities of practice. We will reconsider some of the results of Bolton and Dewatri-point later in the article when we expose ours. Other articles are also dedicated to the search of an efcient hierarchy structure for organizations. They consider the hierarchy as a distribution of authority in the organization (like Hart and Moore, 2005) or as a structure of the communication channels (like Bolton and Dewatripoint, but also Garicano, 2000, Visser, 2000). The rst line of work is quite far from the focus of our article. The second line is more connected to our work, but the general approach and the main point of attack of our article is quite different from this literature where the structure of the network is generally given (it does not result from the learning and the interaction of the agents) and the results depend on the particular cost function used to evaluate the burden of the communications. Moreover these articles are not interested in the quality of the decisions taken using the treated information (except Visser, which we will discuss later in the article). With a very different approach, Delmastro (2002) tries to evaluate the determinants of the hierarchical structure in Italian plants econometrically but neglects many dimensions studied by the theoretical literature, probably because of the difculty of observing these dimensions empirically. The development of ner databases should allow a stronger convergence between the theoretical literature and econometric studies, but our article attacks these questions using a totally different approach. By resorting to computer simulations based on the multi-agent system paradigm, this contribution explores the performance of various organizational settings and, in particular, the role of communities in the performance of a rm. Mainly, two types of structures are contrasted: a pure network of communities of practice and a hierarchy in charge of the division of tasks. Further, within the latter, one can either authorize the potential emergence of communities by allowing the communication between agents or remain in a strictly top-down decision process. It is then possible to compare the outcomes of each of these organizational settings. This work is organized as follows. First the different components of the model (the role and capabilities of individual agents and the overall structures) are presented. Next, the results in terms of comparative performances of organizational structures and the specic impact of learning and communities are considered. The last part of the article uses regression trees to analyze the role of the parameters of the model using the Monte Carlo method. To our knowledge, this is the rst article that uses regression trees for analyzing the results of simulations. Their introduction corresponds to a contribution of this work at a methodological level. The qualitative results obtained using the Monte Carlo method are very robust. 2. The model The model we present here1 is divided in two parts. The rst one (that we call communities of practice and CP hereafter) is a collection of agents having to solve a ow of problems and endowed with the abilities to learn by themselves or by interacting with one another. Each problem is randomly chosen from the problem space and is sent to an agent chosen randomly. The phone calls received by the members of a help desk could correspond to such a process. If the problem does not belong to the realm of specialisation that the agent has built over time through learning

1 The javadoc information about the classes used in our program can be consulted at the following web site: http://beagle.u-bordeaux4.fr/yildi/hiercp/index.html.

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

671

processes, it refuses the problem and passes it to the next agent in a randomly constructed list of agents. Otherwise it accepts the problem and tries to give an answer using its rule set. If it does not possess a satisfactory rule, it consults the community that it has constructed over the past interactions. If the agent does not receive any response from the contacted agents, it passes the problem to the next agent in the random list. If the last agent in the list cannot provide any answer, the problem remains unsolved (the system does not have the competencies to deal with this specic problem). Such a situation corresponds to a loss for the rm and hence a decrease in performance. The overall dynamics thus result both from individual learning by doing and learning by interacting between agents. We will discuss these elements in more detail later in this section. The second part of the model consists of an hierarchical structure where a manager is added to the community of the agents. The role of the manager is to receive the problems from the environment and to select the agent it deems the most able to solve this problem. Its role is thus to carry out a division of labor. We called this part of the model hierarchy with delegation (HD, hereafter) since the manager has to ask its subordinates for advice in order to answer the problems submitted by the environment. In addition, we consider two possibilities: either agents are able to communicate horizontally or not. The article questions the various dimensions of these different organizational settings with a particular interest in the role of communities in the global performance of a rm. In the rest of this section, we present the problem space that organizations are facing, the learning processes of agents, the implemented communication processes and the role of the manager. 2.1. The environment The agents in the model face a continuous ow of classication problems. Such problems are quite common in organizational contexts (cf. for example Carley and Lin, 1997, Ouksel and Vyhmeister, 2000). In this article, the simulation model is inspired by a specic empirical context: a population of agents whose task consists in allocating nancial warranties for bank loans to small enterprises. For each enterprise, a given agent has to determine a certain amount of warranty by considering a set of criteria. In their endeavor to determine the optimum level of warranty, agents can communicate with each other. They can thus exchange the best practices concerning a particular subspace of the problem space. In this way, communities of practice are susceptible to emerge. Formally, we suppose that a project is characterized by two criteria stocked as a eight bits long binary string. The rst four bits code the value of the rst criterion and the four last bits the value of the second one. The problem space is thus a grid, and each binary string represents a particular situation and occupies a position on that grid. An action is associated with each situation (for the agent, the various possibilities take the form [condition:action], where the condition is the binary string and the action is an integer). According to each particular situation, the agent must propose an amount of warranty (the action): for example, 10, 20, 30, 40, 50 or 75% (respectively coded as actions 05; see Fig. 1). The linear nature of the complementarity that exists between these two dimensions of the projects implies smaller surfaces when we get closer to the origin. That could make the classication more difcult, and one of the environments we have tested corresponds to rectangular surfaces. The structure we have retained for this article does not imply any noticeable difculty for learning in comparison with the rectangular one. The linear complementarity relation better corresponds to the nature of the problem we study (riskiness vs return of the projects, for example).

672

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

Fig. 1. The problem space.

The task of the agents is then to determine the amount of warranty associated with each situation. The gap between the correct amount of warranty and the answer proposed by the agent determines the reward of the agent. The reward is computed as follows: f (a) =
e(|a a|)

(1)

where C is a constant, a* is the correct answer expected by the environment and a the answer proposed by the agent. If the agent proposes an amount of warranty lower than the optimum, it does not obtain the full amount of commissions that it could have expected from the situation. If, on the contrary, it proposes a value greater than the optimal one, then in case of bankruptcy of the warranted rm, the loss would be greater than for the correct answer. The reward hence takes into account these two types of costs. The agent uses the reward it receives from the environment to actualize and rene its behavior. 2.2. Modeling of individual learning Each agent has to classify various situations (projects represented as binary strings) in one of the six areas dened by these six possible actions (see Fig. 1). Its experience (the results it obtained from past trials) must help it in rening its judgement. This process corresponds to the individual learning process. This is a typical classication problem, and given the multi-agent system we adopt, a rather natural way to represent this kind of process is the learning classier system (LCS). Indeed, a LCS is a relevant tool for simulating relatively realistic learning processes (see Lanzi and Riolo, 2000 and Holland et al., 1989). More specically, the structure we adopt for each individual agent is an XCS (Wilson, 1995), an algorithm from the family of LCS2 . An XCS receives a signal (problem) from the environment through a detector (Fig. 2). This signal is then compared to the set of rules the agent has in its memory. These rules are made of a condition part and an action part. Moreover, several parameters are associated with each rule: p, the prediction of performance, , the error associated to the prediction
2 A very detailed algorithmic description of this classier system can be found in Butz and Wilson (2000,2001). Our article is not the rst one to use LCS to represent adaptive individual learning. For example, Schulenburg and Ross (2000) uses LCS to study the performance of an articial stock market on real data.

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

673

Fig. 2. Schematic XCS (adapted from Wilson). The Match set is constructed using the rules that contain the signal 0011 in their condition part. The sign # corresponds to a joker that can either match 0 or 1. This possibility is behind the capacity of the Classier System to generalize from observations.

of performance, F, the tness that measures the accuracy of the rules (a function of the inverse of the error) and that is used in the genetic algorithm used by the XCS to explore the problem space. When the genetic algorithm discovers new rules, they are introduced into the population. For each problem, rules regarding which the condition part ts the signal from the environment are stored in an array called the match set. A prediction array is then constituted in order to evaluate the reward associated with each advocated action. This array contains for each possible action the average relative performance (tness prediction) of the rules that prescribe this action in the match set. If no rules in the match set prescribe a particular action, the corresponding value in the prediction array is nil. Based on this array, an action is chosen and triggered in the environment. In our implementation, the agent can select the action predicting the highest reward from the environment or use a roulette wheel process to select the action to be red. A parameter selectAg controls the choice of the selection mode: when set to 1, the best action is chosen; when set to 0, the roulette wheel is used. In the latter case, each rule has a positive probability of being chosen, and this probability increases with the performance of the rule. In our setting these two values of selectAg (0,1) respectively correspond to agents who favour exploration (0: roulette wheel) or exploitation (1: best action). The environment then rewards the XCS that uses this reward to adjust the parameters of its rules. Based on the literature on communities of practice, we endow the agents with the following characteristics. An agent seeks to develop its competencies on a given practice that does not constitute its entire activity but is nonetheless deemed critical by the agent. Moreover, the learning

674

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

process of each agent is oriented: new rules produced by the genetic algorithm are inserted in its memory only if they are in its realm of specialisation. The surface of the specialisation is controlled by the parameter compRange that can rank from 0 to 1. The variable compRange corresponds to the maximal possible distance between the specialisation area of the agent and the submitted problem. For each submitted problem, the agent computes the maximal relative distance3 between the condition part of the best rules in its classier set (the ones with a tness better than the average of its classier set) and the submitted problem. If this distance is higher than compRange, the agent considers that the problem is too far from its specialisation area and passes the problem to the next agent in a randomly composed list of agents. compRange also intervenes in the mutations in order to impede their capacity to disturb the learning process of the agent (the Genetic Algorithm is niche-based in the specialisation area of the agent). When compRange is equal to 1, the agent is not specialised and can accept all the problems submitted by the environment, but when it is close to 0, the agent becomes highly specialised. This is intended to account for the focused learning taking place in communities of practice. An agent has a limited memory that does not permit it to deal with the whole problem space by itself. 2.3. Learning by interacting In our implementation, a parameter, communication, determines whether or not the agents have the ability to communicate. This parameter can take the values 0 or 1; when it is set to 1, the communication is allowed. We now present the communication process (see Fig. 3). The agent can have no rule satisfying the signal from the environment or have only a poor rule (with a prediction of performance below an acceptable threshold, common to all agents). If an agent does not possess a satisfying rule to answer a problem but the problem lies nonetheless in its realm of specialisation, it engages a communication process by requesting help from other agents in its community. Each contacted agent returns the best matching rule it has in its classier set. The requesting agent collects all answers and evaluates them. If a received rule offers a better prediction of performance, then it adopts it. If predictions are equivalent (either several rules it received have identical prediction, or the rules it received and the ones it already possesses have identical prediction), it compares the prediction error and opts for the rule having the lowest one. In case of equivalence between prediction errors, it compares values of tness and chooses the rule with the highest tness. When these parameters are not discriminating, the agent chooses a rule at random. By accepting a rule, the agent copies it in its memory and uses it to answer the signal from the environment. Each agent uses an address book to keep in memory successive communications with different partners. This address book is originally empty. When an agent rst launches a communication process, it asks a percentage (consultRatio) of the whole population at random. It then stores in its address book the references of agents answering him (not only the ones giving him the best answers). In the following periods, the agent will question the agents present in its address book.

3 This distance is normalized by the maximal possible distance in the environment. As a consequence, it belongs to [0,1] as compRange.

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

675

Fig. 3. Solving problems in CP. The agents only accept to handle problems that are close to their specialisation area. If an agent accepts a problem but does not have a solution for this problem, it consults its community (contained in its addressbook). If a member of its community provides it with a satisfactory solution, this solution is integrated in the population of rules of the initial agent (the agent learns this rule).

If none of these agents answer its enquiry, it will then ask once again consultRatio of the total population of agents, which can be seen as an exploration of its social environment and a search for new partners. In order to avoid the presence of non-relevant agents in the address book, a counter (exval) is associated with each agent present in a given address book. If an agent does not answer to exval consecutive enquiries, it is removed from the address book. exval is thus a parameter indicative of the stability of links established by agent and of the general degree of trust in the organization. Hence, in order to enhance its practice, an agent engages with other agents in the building and the sharing of a common repertoire of knowledge resources. In particular, this implies that the agent is always willing to cooperate with other potential members of a community and that there is no direct cost associated with the establishment of a new relation; the agent then nurtures this community of practice with its experience and in turn relies on this community to enhance its practice. Behaviors of agents are then twofold; they are engaged in a practice involving individual learning, and they are committed to social interactions. However, these two parts are interrelated in that communication informs practice and practice fosters exchanges. 2.4. The role of the manager The second organizational structure implemented consists in the simple addition of a manager on top of the structure presented so far. The environment remains the same. However, in this new

676

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

Fig. 4. Problem solving procedure of the manager.

organization, the signals from the environment are directly received by the manager. The duty of the manager is to allocate the task of answering a given signal to the agent most suited for this specic problem. The role of the manager can then be seen as twofold. On the one hand, it has to allocate efciently the tasks as they are submitted by the environment. On the other hand, since the allocation of tasks made by the manager will inuence the learning processes of the agents, it can also be viewed as a guide for the learning of agents. From the modelling standpoint, the manager is an XCS similar to the ones used for modeling agents. The condition part of the managers rules corresponds to the signals arising from the environment. The action parts consist of the references to the agents (i.e. each agent is referenced to by an integer). To choose an agent for answering a given problem, the manager proceeds as follow (see Fig. 4). First, the manager tries to nd an agent by selecting in its memory a rule that matches the signal from the environment. If it does not have any rule, then the manager chooses an agent that is not often called and that has a specialisation realm close to the signal submitted by the environment.4 In the latter case, the manager adds in its memory a new rule of which the condition part is the signal from the environment and whose action part is the reference to the selected agent. As in the case of agents, the manager can select the rule with the best prediction of performance in its action set or use a roulette wheel process to chose the action. This is controlled by the parameter select Man (when select Man is 1, the rule with the best prediction is chosen; when it is 0, the roulette wheel is used). These two values again correspond to a manager who favours exploration (0: roulette wheel) or exploitation (1: best action). Once an agent has been selected by the manager, the problem is passed to this agent and it tries to answer it.

4 The fact that the manager only does one random experiment is not really restrictive. When communication is allowed, several other agents will generally be questioned by the initially chosen agent. We have also done experiments allowing the manager to call some (variable) number of agents as long as it cannot nd an agent that would accept to treat the problem. The results we obtained are very similar to the ones exposed in the next section.

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

677

To answer the problem, the agent can either solve it alone or ask for help from other agents (if communication is set to 1). However, the problem is not passed to another agent when the selected agent cannot answer. In that case, the problem remains unsolved (the organization has not been able to develop a policy for this problem), and incurring a cost that corresponds to a missed opportunity for the organization. The selected agent returns its answer directly to the environment that evaluates it and gives a reward. The reward is then used by the agent and the manager to carry on their learning processes and rene the parameters associated with their rules. 3. Simulation protocol and main results 3.1. The simulation protocol and methodology Given the complexity of the interactions we model and the strong non-linearity of the decision processes of the agents, we adopt a methodology that allows quite a systematic exploration of the parameter space of the model. This methodology is close to the Monte-Carlo method. For each of our two models, we run 500 series of 15,000 problems5 each where the results from each problem has a probability of 1% of being saved. Thus, for each run we obtain an average number of 150 randomly chosen observations for all the measured variables. All series are initialized with a randomly drawn vector of values for the main parameters of the model. As a result, we obtain, for each model, a set of 75,000 observations covering quite a diversied subset of the parameter space and the problem space. The values from which different parameters are drawn can be read in the Appendix A.6 We do not necessarily discuss in the text all the parameters that appear in this appendix but only the most signicant ones. We analyze these random samples using box plots (giving the four quartiles of the distributions of the variables), Student tests between subsets, histograms and regression trees. The statistical analysis is conducted using R (see R. Development Core Team, 2003). The gross performance ( [1000,1000]) is the most direct indicator of the organizational efciency. When the organization faces a problem, it can nd a solution and obtain the actual performance returned by the environment, or it can be unable to provide any solution, in which case the problem remains unsolved. The performance takes into account this missed opportunity; it is computed in this case using the negative of the past average performance (the opportunity that the organization has missed by being unable to solve the problem). This performance criteria is useful for us in order to distinguish rms that generally can solve problems but miss some opportunities from the ones that are never able to solve problems. Missing opportunities can be fatal in a competitive environment as is emphasized by Visser. If the organization is a help desk, not being able to solve the problems of the customers can be very frustrating for them, and they can be lost to the competition. If the organization is the nancial warranty providing rm of our example, projects that are not handled can be warranted by the competition and the rm can loose its credibility and its market share. Given that we desire to evaluate the capability of the organization to handle problems as well as the quality of this treatment, we must qualify this capability. The relative weight of negative performances in comparison with the positive ones nicely indicates the capacity of each organization to treat problems.
5 15000 problems correspond to the yearly average number of cases considered by the renancing company we have analysed. 6 This appendix is downloadable from our web site: http://beagle.u-bordeaux4.fr/yildi/les/appendixhiercp.pdf.

678

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

However, the gross performance neglects another important dimension of the efciency of the organizations: the cost of communication. If a solution is found after a tremendous amount of communication between agents, it would correspond to a relatively high organizational cost. In order to take into account the cost as well as the benet of each decision, we also analyze the total number of communications in each case. 3.2. Main results of the simulations We focus the analysis of the results on issues concerning the performance of organizations in problem solving. We rst proceed by comparing these organizational forms from the point of view of gross performance. These rst results are rened afterwards through the analysis of the role of communications and communities of practice. The last step of the analysis claries the role of different parameters in these results using regression trees. 3.2.1. Comparative analysis of the organizational performance We give in Fig. 5 the comparison of the distributions of gross performance for all cases in both organizational forms. These boxplots show the four quartiles of these distributions: the statistically signicant minimum and the maximum correspond to the extreme end of the whiskers, while Q1 and Q3 correspond to the edges of the central box and the median corresponds to the horizontal line inside the box. We observe that communities of practice (CP) is globally able to treat more problems than the hierarchy (HD) (hence it better avoids negative performances) and it treats them better than HD since the highest performances have a higher frequency in CP. The Student test results provided below conrm this result and show that the average of gross performance in CP is higher than the one in HD. This strong result calls for a more detailed analysis. As will clearly appear in the rest of this analysis, one of the phenomena that plays a crucial role in the performance of these organizational structures is the possibility of communication. Fig. 6 compares the distribution of the gross performances of HD and of CP between two cases:

Fig. 5. Comparison of gross performance between organizations.

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

679

Fig. 6. Comparison of gross performance between organizations with (1) and without (0) communication.

with communications (1) and without communications (0). The graphic (a) shows that the communities of practice need communications to better avoid the negative payoffs (due to unsolved problems), while the hierarchy needs communications for attaining better payoffs (graphic (b)). As a consequence, communications do not play the same role in these two organizational structures. The comparison of results between HD and CP shows that CP is able to attain, on average, better payoffs than HD. The organizational structure plays the main role in the determination of the performance, but communities with communication (CP (communication = 1)) attain the highest performances and the complete ordering between these four cases is given by CP(1) CP(0) HD(1) HD(0). (2)

Student tests between different cases concerning the gross performance give the following results: H1: CP < HD t = 64.2951, df = 149390.3, p value = 1 H1: CP (1) < HD(1) t = 51.3433, df = 122334.1, p value = 1 H1: CP (1) < CP (0) t = 8.9367, df = 31741.86, p value = 1 H1: CP (0) < HD(1) t = 27.6562, df = 32190.58, p value = 1 H1: HD(1) < HD(0) t = 41.9592, df = 12777.20, p value = 1 H1: CP (0) < HD(0) t = 55.6133, df = 18648.14, p value = 1

In Fig. 7, we consider the total number of communications in CP and HD. It clearly shows that the number of communications is the highest in CP. Moreover, Fig. 6 evidences that com-

680

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

Fig. 7. Comparison of the number of communications.

munications signicantly enhance the gross performance of the organizations but they could also be seen as an important source of organizational cost (Ishida and Ohta, 2001; Butler, 2000). However, this cost cannot be specied without introducing an ad hoc bias. Nonetheless, if this cost is really consequential, it could eliminate the contribution of communications to gross performance. Organizations are thus facing a trade-off, having to nd a balance between the number of communications that would increase their performance and a reasonable cost of these communications. In the rest of the article, we will focus the analysis on the settings where communication is allowed. 3.2.2. Emergence and performance of the communities of practice The comparative analysis of performance indicates that communications among agents, outside the hierarchical line, play a role in the behavior of the system as a whole. In our setting, to allow communication is to allow agents to build relational structures beyond the hierarchy. In other words, in our setting, when communications occur, one can hope to observe the emergence of communities of practice. To account for the emergence of their communities, we here use an indicator well-known in social network analysis, cliquishness (Newman et al., 2001; Watts, 1999; Wasserman and Faust, 1994), which captures the idea that two agents connected to another agent are likely to be connected to one another. It is thus an indicator of the existence of groups of tightly connected agents within the graph. Given the assumptions we adopt about the agents and their behaviors, such groups would correspond here to communities of practice. In a directed graph, cliquishness represents, for a vertex with n neighbors, the fraction of potential edges that actually exist between them: the full connection of all the n neighbors with each other would correspond to n(n 1). Cliquishness is the average of this measure over all vertices. If G(I, d) is a graph, with I being the set of the n nodes of G and d being its adjacency matrix, the average cliquishness of G can be dened using the following notation. If d(i,j) is the distance between the vertices i and j, Vi = {j|d(i,j) = 1} is the direct neighborhood of i. Let ni = #Vi

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

681

and mi =
j,l

d (j, l)j, l Vi such as d(j,l) = 1. The cliquishness of G can be computed as 1 n mi . ni (ni 1)

C=

(3)

Cliquishness can consequently rank from 0 to 1. When it is equal to 0, no triangles exist within the graph. When cliquishness equals 1, the graph is made of fully connected subgraphs. Given that communications in our model create new links, as well as eliminate the existing ones that become irrelevant, the passage of time alone does not automatically increase cliquishness. Fig. 8 shows that communities do not emerge solely in the CP model, but also in HD. Nonetheless, the highest values of cliquishness are only reached in CP. As a consequence, we observe in the latter either a higher number of communities or more tightly knit ones. A potentially interesting feature of communities is their contribution to the attainment of high performances. In Dupou et et al. (2003), we characterize the emergence of communities of practice in the CP model. To do so, we use three indicators (betweenness, degrees and cliquishness). We show that full-edged communities of practice only emerge for certain values of these indicators. These congurations correspond to values of cliquishness between 0.2 and 0.3. In order to study the role of communities in the overall performance, we consider the distribution of cliquishness in the cases corresponding to the maximal performance. Fig. 9 shows that in CP as well as in HD, the most frequent values of cliquishness, when the performance is maximal, are the ones that also correspond to the emergence of the communities of practice (the distributions possess a cumulation around 0.20.3). These observations will be conrmed below in Figs. (12a) and (13a). These results show that for both types of organizations, the highest performances are attained mainly in the cases where the communities of practice emerge. Even if it is practically quite arbitrary to determine a limit value of cliquishness corresponding to the emergence of communities of practice, the correspondence between the highest performances and the presence of structures similar to communities of practice is quite strong.

Fig. 8. Comparison of Cliquishness.

682

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

Fig. 9. Distribution of cliquishness for the cases with the maximal performance.

Our results have so far focused on the most striking results of the model. However, the understanding of the mechanisms behind these results requires further explorations. 3.2.3. Behind the curtains: the role of the main parameters We analyze the role of different parameters of the model using regression trees (Venables and Ripley, 1999, chapter 10). A regression tree establishes a hierarchy between independent variables using their contributions to the overall t (R2 ) of the regression. More exactly, it splits the set of observations in sub-classes characterized by their value in terms of their contribution to the overall t and of their predictions for the dependent variables (all parameters that are modied by the Monte Carlo procedure are included as explanatory variables in each of the following regressions). This value is validated against a fraction (10%) of the sample that is not used during the estimation. Regression trees are very exible and powerful in the clarication of the structure of the observations. We give in the next paragraph a step-by-step interpretation of the regression tree exposed in Fig. 10. We now consider the determinants of the main performance variables of our model.

Fig. 10. Regression tree for the gross performance in HD.

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

683

Fig. 11. Regression tree for the gross performance in CP.

3.2.3.1. Determinants of the gross performance. The Figs. 10 and 11 exhibit the variables that contribute more than 1% to the overall t in both cases (HD and CP) with communications. In Fig. 10, the rst condition (selectMan < 0.5) divides the observations in two subsets. On the left, we have the observations for which this condition is fullled, and on the right, the observations that correspond to (selectMan 0.5). This is a common feature of all trees; on the left of a condition we have observations for which the condition is true, and on the right, the ones for which it is false. The observations on the left, corresponding to (selectMan < 0.5), are again separated in two sub-groups by another condition concerning the specialization of the agents: (comprange < 0.5787). The nal nodes of the tree indicate the number n of observations concerned by these conditions and the expected value of the dependent variable (here the gross performance) in this sample. For example, (selectMan < 0.5) and (comprange < 0.5787) are fullled for a set of 20307 observations (over a total of 20307 + 13454 + 5661 + 12875 + 13594 = 65891) in which the expected gross performance is 243.2. At the other side of the last condition, we have 13454 observations verifying (selectMan < 0.5) and (comprange 0.5787). The condition (selectMan < 0.5) in fact corresponds to selectMan = 0; the manager favours exploration over exploitation. In this case, positive performances (169.2) can only be expected if the agents are not too specialized (comprange 0.5787). The rest of the tree (and all other trees) can be read following the same scheme. HD with communications In the case of the hierarchy, the main determinants of the gross performance are selectMan, compRange and selectAg. The other variables play only a marginal role. Higher performance is obtained when selectMan = 1 and, hence, when the manager chooses the best agent in its agents set for each problem (instead of choosing each agent with a positive probability, even if it is increasing with the performance, selectMan = 0). When the manager favours exploitation (it uses the best action rule), the agents should not be too specialized (compRange 0.20) in order to have an interesting global performance. This performance is even higher if the agents also favour

684

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

Fig. 12. Determinants of gross performance in HD. (a) Role of Cliquishness (b) Role of comprange (c) Role of the behavioral selectors.

exploitation (they use the best action rule (selectAg = 1) for choosing their actions). When the manager uses a more explorative strategy (selectMan = 0), the organization is only able to treat problems if the agents are loosely specialized [Fig. 10]. CP with communications In the communities, the hierarchy between the determinants is slightly different. Even if selectAg plays the main role in the segregation of different cases, the specialisation of the agents clearly appears as the main determinant of the performances. The community can only attain signicant performances if the agents are not too specialised (compRange 0.13). When conditions upon compRange and selectAg are fullled, then time matters (since the variable period correspond to the number of the submitted problems). There is a learning trajectory of the system over time. Since time does not play this privileged role in HD, it seems that the initial conditions in CP are less determining than in HD. It is interesting to observe that other dimensions of the communication (like the consultRatio) play only a marginal role in the determination of the gross performance. Their role only appears when we consider more qualitative aspects of the performance (Fig. 11). The boxplots of Figs. 12 and 13 give the distribution of gross performance as a function of another variable (for example, cliquishness). These values are split in two groups, on the left we have the values of the performance when the explicative variable is lower than its second quartile (its median, Q2 ) and on the right, we have the complementary cases where the variable is higher than its median.

Fig. 13. Determinants of gross performance in CP. (a) Role of Cliquishness (b) Role of comprange (c) Role of selectAg.

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

685

The comparison of Fig. 12 with Fig. 13 reveals an interesting structural regularity: higher performance is associated with higher cliquishness and lower specialisation (higher comprange), and exploitation allows the hierarchy to increase its performance while permitting the communities to reduce negative performances. This again conrms the positive relationship between the emergence of the communities of practice and the global performance and explains how this mechanism works: higher cliquishness increases the capacity of CP to cover the problem space, while it increases the accuracy of problem solving by HD. We should not forget that even if the manager and the agents favour exploitation, some minimal amount of exploration is always present through the genetic algorithm. 3.2.3.2. Determinants of communications. The number of communications in the organization will generally imply a cost as we have stressed above (Ishida and Ohta, 2001; Butler, 2000). Whatever the nature of this cost (linear, non-linear, etc.) it will generally be increasing with the number of communications. It is legitimate then to ask if it is possible to limit the communications when their cost is prohibitive. We consider the determinants of the communications in Fig. 14 in order to identify the parameters that the organization could weight in order to reduce this cost. HD with communications In HD the number of communications remains relatively limited. It is sufcient to limit the possibility for the agents to consult other members in order to minimize these costs. This is not a tautology, since the actual number of communications also depends on other variables such as the specialization of the agents. Even if strongly restricting the possibility of consultation is not possible, a manager who strongly favours exploitation over exploration in its decisions will also be able to limit these costs. Last but not least, an intermediate level of specialisation (compRange [0.17,0.63]) of the agents can also be used to contain these costs (Fig. 14a). CP with communications In CP, one must denitely avoid a very high specialisation (compRange < 0.1) for the agents, but this is not enough. Some limitation of the possibilities of consulting other agents is necessary

Fig. 14. Determinants of the number of communications.

686

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

to limit the communication costs as in HD (consultratio < 0.43). When this is again not possible, costs can be contained if agents are incited to favour exploitation (selectAg = 1) (Fig. 14b). We nd here the trade-off emphasized by Bolton and Dewatripoint between specialisation and the cost of the communications, especially in CP (lower compRange corresponds to a considerably higher number of communications). Moreover, given that we take into account the quality of the treatment of the problems by the organization (a dimension neglected in Bolton and Dewatripoint) the problem of the organization is truly a trade-off; since the limitation of the communications will generally have a negative impact on gross performance. On the other hand, given that consultRatio does not strongly condition the gross performance (it does not appear in Figs. 10 and 11), it is possible to overcome this trade-off: the highest performances are attainable even if we have consultRatio < 0.3 in HD and not too specialized agents in CP (compRange > 0.1) with a reasonable weak consultRatio (<0.43). As a consequence, the global efciency can be reinforced by crossing the prescriptions of Figs. 10 and 11 and the ones of Fig. 14. We also remark that without such ne-tuning, HD can successfully assure such a trade-off, given the very low level of communications in this structure and the reasonable level of the performances. CP can assure more systematically positive performances, but it can also incur a tremendously higher level of communications. 3.2.3.3. Determinants of cliquishness. The role of cliquishness in the characterization of the emergence of communities of practice has been studied by us in a previous article (see Dupou et et al.) and above. In particular, we have established in this article that the communities of practice should correspond to a cliquishness around 0.3. Figs. 15 and 16 show that the determinants of the social structure are slightly different in these two organizational structures.

Fig. 15. Regression tree for cliquishness in HD.

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

687

Fig. 16. Regression tree for cliquishness in CP.

HD with communications Three variables intervene in the emergence of communities in the case HD: consultRatio, exval and compRange. When consultRatio is low (consultRatio 0.11), then no communities can emerge and the value of cliquishness is almost one of a random graph. When 0.11 consultRatio < 0.35, then exval has to be quite high, that is, there must be some kind of delity or trust in relationships between agents in order to observe the emergence of communities. Even when consultRatio is higher, exval still has to be above 3.5 to obtain the highest values of cliquishness. In any case, compRange has to be above a certain threshold to have values of cliquishness corresponding to full-edged communities of practice (Fig. 15). CP with communications The same three parameters are the determinants in the emergence of communities in the case CP. However, compRange (that is the level of specialisation of agents) plays here a less important role than in the case HD. As in the previous case, when consult Ratio is low, high values of exval help reaching reasonably high values of cliquishness (i.e. exval 9.5 implies a cliquishness of 0.14). The highest values of cliquishness are attained for high values of consultRatio and exval. Hence, in order to build communities of practice, agents need both a great ability to screen their social environment and a certain stability in their relationships (Fig. 16). The main determinant of the importance of communities of practice is therefore the possibility of communications between agents. This is of course not a very surprising result in itself, but

688

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

given that the conditions for it are quite demanding (consultRatio > 0.35), communications are not enough; some degree of trust in the organization is also necessary, and communities of practice are not a simple epiphenomenon.

4. Conclusion This article analyzes the potential impact that informal social structures, communities of practice, can have upon the performance of rms or subunits of rms. The rst result of this work is to show that community structures are efcient for competence building, particularly if one considers learning in the long term (i.e. without strong time constraints). However, real rms are required to answer demands from the market in a timely fashion. In these cases, this work tends to show that hierarchy altogether with communities of practice is the most efcient structure among those we explore. This paper thus backs the claim made by Bowles and Gintis (2002) that hierarchy and communities are complementary rather than substitute modes of governance. When the cost of communications are important, HD can successfully assure a satisfactory trade-off; it implies a very low level of communications and a reasonable level of performance. When the reliability of the performance is the issue, CP can assure more systematically positive performances, but it can also incur a tremendously higher level of communications. Moreover, the examination of the interrelations between variables reveals that two of them are especially crucial for the emergence of communities and their performances: the degree of specialisation of agents and the conditions of communication between them. If the former clearly depends on the individual agent, the second can be seen as part of the environment in which communities exist (e.g. the efforts made by the management to ease communication, the existence of an intranet, etc.). In the case of HD, one must add the behavior of management to the two previous parameters. The way managers apprehend their task has an inuence both on the conditions of emergence of communities within the rm and on the global performance of the organization. It is the interplay between agents idiosyncratic capabilities and managers behavior that ends up in an organizations specic outcome. However, we are well aware that this can only constitute an exploratory work on these issues. Regarding communities, the detailed dynamics of their behavior has not been explored. In particular, the legitimate peripheral participation process that Lave and Wenger put at the heart of the evolution of communities structure has not been explored. Also, an interesting development of the model could follow Hart and Moore and allow for a differentiated specialization of the agents, some of them being less specialized than others. The particular role that such agents could play in the network can be very interesting to study. Enriching the possible structures for the hierarchy could also be very interesting since this would allow a better comparison of our model with the hierarchy models developed by Bolton and Dewatripoint or by Garicano. We should not forget that the structure of the communication hierarchy is endogenous in our model and its structure could be more directly compared with the structures considered by Bolton and Dewatripoint (conveyor belt or strict hierarchy, for example). In an article more directly focused on the hierarchy (instead of the communities of practice), some indicators from the social network analysis literature could be used to characterize the hierarchical structure of the agents communication network. Besides, concerning the incentives and motivations of agents to enter in a community, we have assumed that the agents are always willing to enhance their individual competencies by resorting to communities. This assumption

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

689

is common in the literature about communities of practice but would certainly deserve a deeper exploration. Our model with hierarchy also focuses on some key aspects and leaves many other features unaddressed. For instance, the fostering of communities of practice is not considered as a strategic issue by the manager, whereas this is certainly a main concern in management studies today. Moreover, we here only dealt with a specic form of hierarchy where the manager is engaged in a learning process. It could also be interesting to investigate a more rigid hierarchy where the manager strongly prescribes the tasks to agents without seeking to take into account their learning curves or specialisation (as, for example, in Carley and Lin). More complex problem spaces (as in Rivkin and Siggelkow, 2003) should also be analyzed since they can condition the learning process of the agents. All these limitations highlight the work that remains to be done in modeling the behavior and the role of communities within rms. It is hoped that this work contributes in opening the way to this future research. Acknowledgements We gratefully acknowledge the comments and suggestions of the participants of the SCE Workshop on Complex Behavior in Economics and Finance 2003 and of the WEHIA 2003 Workshop. We are very grateful to two anonymous referees who have considerably contributed to the readability of this article. Financial support of the CCRRDT program of R egion Aquitaine is gratefully acknowledged by Murat Yildizo glu. References
Amin, A., Cohendet, P., 2003. Knowledge Practices, Communities and Competences in Firms. Oxford University Press, London. Bolton, P., Dewatripoint, M., 1994. The rm as a communication network. The Quarterly Journal of Economics 109, 809839. Bowles, S., Gintis, H., 2002. Social capital and community governance. The Economic Journal 112, 419437. Brown, J.S., Duguid, P., 1991. Organizational learning and communities of practice: toward a unied view of working, learning and innovation. Organization Science 2, 4057. Brown, J.S., Duguid, P., 1998. Organizing knowledge. California Management Review 40, 90111. Butler, B.S., 2000. Communication cost, belief development and structural change: a dynamic model of networked communities of practice. CASOS working paper. http://www.casos.ece.cmu.edu/. Butz, M.V., Wilson, S.W., 2001. An algorithmic description of XCS. In: Lanzi, P.L., Stolzmann, W., Wilson, S.W. (Eds.), Advances in Learning Classier Systems, Third International Workshop (IWLCS-2000), Lecture Notes in Articial Intelligence (LNAI-1996). Springer-Verlag, Berlin. Carley, K.M., Lin, Z., 1997. A theoretical study of organizational performance under information distortion. Management Science 43, 976997. Crozier, M., Friedberg, E., 1977. Lacteur et le syst eme. Points Seuil, Paris. Delmastro, M., 2002. The determinants of the management hierarchy: evidence from italian plants. International Journal of Industrial Organization 20, 119137. Dupou et, O., Yildizo glu, M., Cohendet, P., 2003. Morphogen ese de communautes de pratique. Revue dEconomie Industrielle 103, 91110. Garicano, L., 2000. Hierarchies and the organization of knowledge in production. The Journal of Political Economy 108, 874904. Hart, O., Moore, J., 2005. On the design of hierarchies: coordination versus specialization. Journal of Political Economy 113, 675702. Holland, J.H., Holyoak, K.J., Thagard, P.R., 1989. Induction. Processes of inference, learning and discovery. MIT Press, Cambridge.

690

O. Dupou et, M. Yldzo glu / J. of Economic Behavior & Org. 61 (2006) 668690

Huberman, B.A., Hogg, T., 1995. Communities of practice: performance and evolution. Computational & Mathematical Organization Theory 1 (1), 7392. Ishida, K., Ohta, T., 2001. On a mathematical comparison between hierarchy and network with a classication of coordination structures. Computational & Mathematical Organization Theory 7, 311330. Lanzi, P.L., Riolo, R., 2000. A roadmap to the last decade of learning classier system research. In: Lanzi, P.L., Stolzman, W., Wilson, S.W. (Eds.), Learning classier system, from foundation to application. Springer-Verlag, Berlin, pp. 3363. Lave, J., Wenger, E.C., 1990. Situated learning: legitimate peripheral participation. Cambridge University Press, New York. Newman, M.E.J., Strogatz, S.H., Watts, D.J., 2001. Random graphs with arbitrary degree distributions and their applications. Physical Review E 64, 026118. Ouksel, A., Vyhmeister, R., 2000. Performance of organizational design models and their impact on organization learning. Computational & Mathematical Organization Theory 6, 395410. Orr, J., 1990. Talking about machines: An ethnography of a modern job. Cornell University Press, Ithaca, NY. R. Development Core Team, 2003. R: a language and environment for statistical computing. R foundation for statistical computing. Vienna: http://www.r-project.org/. Rivkin, J.W., Siggelkow, N., 2003. Balancing search and stability: interdependencies among elements of organizational design. Management Science 49, 290312. Schulenburg, S., Ross, P., 2000. An adaptive agent based economic model. In: Lanzi, P.L., Stolzman, W., Wilson, S.W. (Eds.), Learning classier system, from foundation to application. Springer-Verlag, Berlin, pp. 263283. Venables, W.N., Ripley, B.D., 1999. Modern applied statistics with S-PLUS, 3rd ed. Springer, New York. Visser, B., 2000. Organizational communication structure and performance. Journal of Economic Behavior and Organization 42, 231252. Wasserman, S., Faust, K., 1994. Social network analysis: methods and applications. Cambridge University Press, Cambridge. Watts, D.J., 1999. Small worlds. Princeton University Press. Wenger, E., 1998. Communities of practice: learning, meaning and identity. Cambridge University Press, Cambridge. Wilson, S.W., 1995. Classier tness based on accuracy. Evolutionary Computation 3, 149175.